Harmonizing Base and Novel Classes: A Class-Contrastive Approach for Generalized Few-Shot Segmentation

Current methods for few-shot segmentation (FSSeg) have mainly focused onimproving the performance of novel classes while neglecting the performance ofbase classes. To overcome this limitation, the task of generalized few-shotsemantic segmentation (GFSSeg) has been introduced, aiming to predictsegmentation masks for both base and novel classes. However, the currentprototype-based methods do not explicitly consider the relationship betweenbase and novel classes when updating prototypes, leading to a limitedperformance in identifying true categories. To address this challenge, wepropose a class contrastive loss and a class relationship loss to regulateprototype updates and encourage a large distance between prototypes fromdifferent classes, thus distinguishing the classes from each other whilemaintaining the performance of the base classes. Our proposed approach achievesnew state-of-the-art performance for the generalized few-shot segmentation taskon PASCAL VOC and MS COCO datasets.